Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR

Page created by Clyde Page
 
CONTINUE READING
Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR
2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)
                                                                                                                                                                 October 25-29, 2020, Las Vegas, NV, USA (Virtual)

                                                                                                                                                                     Go-CHART: A miniature remotely accessible self-driving car robot
                                                                                                                                                                                                             Shenbagaraj Kannapiran and Spring Berman
2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) | 978-1-7281-6212-6/20/$31.00 ©2020 IEEE | DOI: 10.1109/IROS45743.2020.9341770

                                                                                                                                                                    Abstract— The Go-CHART is a four-wheel, skid-steer robot
                                                                                                                                                                 that resembles a 1:28 scale standard commercial sedan. It is
                                                                                                                                                                 equipped with an onboard sensor suite and both onboard and
                                                                                                                                                                 external computers that replicate many of the sensing and
                                                                                                                                                                 computation capabilities of a full-size autonomous vehicle. The
                                                                                                                                                                 Go-CHART can autonomously navigate a small-scale traffic
                                                                                                                                                                 testbed, responding to its sensor input with programmed con-
                                                                                                                                                                 trollers. Alternatively, it can be remotely driven by a user who
                                                                                                                                                                 views the testbed through the robot’s four camera feeds, which
                                                                                                                                                                 facilitates safe, controlled experiments on driver interactions
                                                                                                                                                                 with driverless vehicles. We demonstrate the Go-CHART’s
                                                                                                                                                                 ability to perform lane tracking and detection of traffic signs,
                                                                                                                                                                 traffic signals, and other Go-CHARTs in real-time, utilizing
                                                                                                                                                                 an external GPU that runs computationally intensive computer
                                                                                                                                                                 vision and deep learning algorithms.

                                                                                                                                                                                        I. I NTRODUCTION
                                                                                                                                                                    In this paper, we propose a miniature mobile robot that can
                                                                                                                                                                 emulate sensing and computation capabilities of a full-size                     Fig. 1.   Small-scale driving testbed and a Go-CHART robot (inset).
                                                                                                                                                                 autonomous vehicle, with the aid of an external GPU for
                                                                                                                                                                 computationally intensive tasks. The Go-CHART can also
                                                                                                                                                                 be driven remotely by a user with first-person views of the                  modular and can be upgraded with updated microcontrollers
                                                                                                                                                                 robot’s environment through video feeds from its cameras.                    and microprocessors, as well as stereo cameras. It utilizes a
                                                                                                                                                                 The Go-CHART can be used to conduct studies on conditions                    system architecture that supports an external GPU capable
                                                                                                                                                                 and challenges that are likely to be experienced by real                     of handling computationally intensive deep learning and
                                                                                                                                                                 autonomous vehicles, but are too difficult to realistically                  machine learning algorithms. In autonomous mode, the Go-
                                                                                                                                                                 replicate in a driving simulator (e.g., crash scenarios with                 CHART is controlled by on-board Raspberry Pi computers
                                                                                                                                                                 complex vehicle dynamics and impacts on surroundings,                        and Teensy microcontrollers, and external computers that
                                                                                                                                                                 dynamic lighting and weather conditions) and too risky to                    replicate many of the sensing and computation capabilities of
                                                                                                                                                                 physically implement at full scale. The Go-CHART can be                      a full-size AV; for example, lane tracking and object detec-
                                                                                                                                                                 used to investigate interactions between human drivers and                   tion using deep learning algorithms. We have also developed
                                                                                                                                                                 driverless vehicles; perform studies on cooperative driving                  a graphical user interface that enables a user to remotely
                                                                                                                                                                 strategies; and test computer vision, deep learning, and                     drive the Go-CHART using a steering wheel and responsive
                                                                                                                                                                 control algorithms for autonomous vehicles. It can also be                   pedals. Additionally, we have developed a small-scale testbed
                                                                                                                                                                 used to collect a wide range of data sets in scenarios that are              that resembles a U.S. driving environment, including roads,
                                                                                                                                                                 difficult to reproduce on full-size vehicles for training deep               traffic lights, signs, and other miniature scenery such as trees,
                                                                                                                                                                 learning algorithms.                                                         which not only increase the realism of the testbed for human-
                                                                                                                                                                    Deep learning and computer vision play a crucial role                     robot interaction experiments, but also provide feature points
                                                                                                                                                                 in controlling autonomous vehicles and are essential for                     in the environment for performing experiments on monoc-
                                                                                                                                                                 the vehicle to exhibit higher levels of autonomy. However,                   ular simultaneous localization and mapping (SLAM). In
                                                                                                                                                                 existing miniature driving testbeds lack the capabilities to                 this paper, we experimentally demonstrate the lane-tracking,
                                                                                                                                                                 support an external GPU. In this work, our contributions                     object detection, and decision-making capabilities of the Go-
                                                                                                                                                                 include the design and development of the Go-CHART, a                        CHART on this testbed.
                                                                                                                                                                 small four-wheel skid-steer robot that resembles a standard                                       II. E XISTING P LATFORMS
                                                                                                                                                                 commercial sedan. It is equipped with an ultra-wide night
                                                                                                                                                                                                                                                 The recently accelerated development of self-driving cars
                                                                                                                                                                 vision front camera, side-view and rear-view cameras, an ul-
                                                                                                                                                                                                                                              has motivated the need to validate controllers for these vehi-
                                                                                                                                                                 trasonic sensor, and a 2D LiDAR sensor. The Go-CHART is
                                                                                                                                                                                                                                              cles on experimental testbeds. Several full-scale testbeds with
                                                                                                                                                                    This work was supported by the ASU Global Security Initiative.            5G-enabled infrastructure have been implemented for testing
                                                                                                                                                                    Shenbagaraj Kannapiran and Spring Berman are with the School              connected and autonomous vehicles, including University
                                                                                                                                                                 for Engineering of Matter, Transport and Energy, Arizona State
                                                                                                                                                                 University (ASU), Tempe, AZ 85281, USA {shenbagaraj,                         of Michigan’s Mcity [18], South Korea’s K-City [19], and
                                                                                                                                                                 spring.berman}@asu.edu                                                       UK’s Millbrook Proving Ground [20]. In addition, Google

                                                                                                                                                                 978-1-7281-6212-6/20/$31.00 ©2020 IEEE                                  2265

                                                                                                                                                                            Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR
TABLE I
has created a miniature Google Street View in tiny realistic
                                                                                         I NTERNAL C OMPONENTS OF G O -CHART ROBOT
reproductions of various international sites and cities [17].
   The paper [1] includes a list of robots costing less than                        Components                            Number       Total Cost
US $300 that were developed within the last 10 years from                                                                 of Units     (US$)
                                                                                    Raspberry Pi 3B                         1          32
the date of its publication. The work proposed a small                              Raspberry Pi Zero                       3          15
educational robot that can drive autonomously through a                             RPLIDAR A1 2D LiDAR                     1          99
traffic testbed. However, the robot lacks the sensors and                           Raspberry Pi camera                     3          15
                                                                                    Raspberry Pi ultra-wide angle           1          11
processing capabilities that facilitate the implementation of                       night vision camera
decentralized robot controllers and computation-intensive                           Teensy 3.2                              1          20
deep learning algorithms.                                                           TB6612 Adafruit motor driver            2          10
                                                                                    Polulu mini motors with encoders        4          36
   The University of Delaware Scaled Smart City [8] is a                            SD cards (16GB)                         4          16
small-scale testbed that has been used to replicate real-life                       Miscellaneous components                -          10
traffic scenarios and optimize control strategies on mobile
systems. The testbed uses a Vicon motion capture system                                                       TABLE II
to localize and coordinate 35 vehicles simultaneously in a                               E XTERNAL C OMPONENTS OF G O -CHART ROBOT
centralized manner.
                                                                                    Components                            Number       Total Cost
   The paper [3] proposes a traffic testbed with a fleet of                                                               of Units     (US$)
miniature vehicles with Ackermann steering. The testbed                             External GPU (1 per robot)              1          500
supports cooperative driving studies and multi-vehicle navi-                        LAN cables (2 ft) (1 per robot)         1            3
gation and trajectory planning. This paper includes a list of                       Miscellaneous components (per           -           10
                                                                                    robot)
miniature robots that also use Ackermann steering. However,                         Dual-band router (1 per ∼60             1          200
the robotic platform used in [3] does not incorporate vision                        robots)
or range sensors and lacks the processing capacity to run
computationally intensive algorithms.
   Several open-source and commercially available miniature                  the external hardware components of the Go-CHART, which
autonomous vehicle robots with onboard GPUs and vision-                      cost approximately US $513 in total (excluding the network
based control have been developed in recent years. The                       router, which can be used with multiple robots).
MIT Racecar [5], MuSHR [21], and Donkey Car [23] are                            The body of the Go-CHART is a 1:28 model of a
open-source self-driving race car platforms with a range of                  standard commercial sedan and features a modular design,
onboard processors and sensors. The AWS DeepRacer [22]                       allowing for easy replacement and addition of components
and NVIDIA Jetbot [24] are commercially available robots                     as necessary. The main body, labeled 3 in Fig. 2, is a single
with Ackermann and differential-drive steering, respectively.                3D-printed piece, while the remaining parts of the frame are
However, all of these platforms lack 360◦ camera views, and                  made from laser cut acrylic.
their restricted onboard processing capabilities preclude the                   Custom printed circuit boards (PCBs) were designed in
implementation of complex deep learning techniques.                          order to densely populate all the circuitry within the limited
   In [9], it was demonstrated that small differential-drive                 volume of the Go-CHART. The PCBs provide several voltage
robots [2] can exhibit certain autonomous functions on                       levels (3.3V, 5V, and 12V) that can accommodate a wide
a driving testbed such as lane tracking and traffic light                    variety of sensors, microcontrollers, and microprocessors.
detection. However, the limited onboard processing power                     The PCBs have open ports for additional sensors, motor
of the robot caused delays in image processing, resulting in                 drivers, and microcontrollers as needed. Four custom PCBs
errors in navigation. Other differential-drive robots such as                were fabricated: (1) a power supply board, which includes
[6], [7] can also emulate particular functions of autonomous                 the voltage regulators and the microcontroller board; (2) a
vehicles, but lack the processing power necessary for real-                  motor driver board, which includes the motor drivers and
time navigation and object detection.                                        all sensors; and (3) front and (4) back LED boards, which
      III. G O -CHART D ESIGN AND C APABILITIES                              control the headlights, tail lights, and turn indicator lights.
A. Mechanical design and circuit boards
                                                                             B. Drive mechanism
   Figures 2 and 3 show SolidWorks renderings of the Go-
CHART, with its components labeled in the exploded view                         The Go-CHART uses a four-wheel skid steering system,
in Fig. 2. The Go-CHART is built from both off-the-shelf                     which is powered by four standard micro metal gear motors
and custom-built components such as custom PCBs, 3D                          with a 51.45:1 gear ratio. The motors drive 32-mm-diameter
printed parts, and laser cut acrylic parts. As described in                  wheels and have Hall effect encoders with a linear resolution
Section III-D, we use the GPU on the NVIDIA Jetson                           of 0.163 mm/tick, enabling the Go-CHART to move at
TX2 module, which costs around US $500. However, the                         speeds between 4 cm/s and 42 cm/s. Two Adafruit TB6612
choice of GPU is flexible, depending on the application.                     motor drivers control each motor separately, using a PD
Table I lists the primary internal components of the robot                   controller to limit the overshoot and dampen the oscillations
and their costs, which total to about US $264. Table II lists                in the robot’s heading angle.

                                                                        2266

           Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR
Fig. 2.   Exploded SolidWorks rendering of Go-CHART, with components labeled.

                                                                               Robot Operating System (ROS) and the OpenCV library [16]
                                                                               installed. The RPi Zero boards run the Raspbian OS and
                                                                               are upgraded with external dual band USB wireless network
                                                                               adapters, soldered directly to the motherboard, to enhance
                                                                               video streaming performance. Low-level control of actuation
                                                                               and sensor data processing are performed by a Teensy 3.2
                                                                               microcontroller (MK20DX256VLH7 processor with ARM
                                                                               Cortex-M4 core running at 96 MHz with 64kB RAM), which
                                                                               can accommodate a large number of PWM and digital pins.
                                                                                  In addition to these onboard components, an external GPU
                                                                               is used to train and implement the neural network models for
                                                                               the deep learning algorithms described in Section V. The
Fig. 3. SolidWorks renderings of isometric view, top view, and front view      external GPU’s capacity can be varied as needed, based on
of Go-CHART.                                                                   the testing requirements. We used a GeForce RTX 2080 Ti
                                                                               GPU to train the neural network models. To keep the system
                                                                               low-cost and portable, we used a low-power NVIDIA Jetson
C. Power supply
                                                                               TX2 module with 256-core PascalTM GPU architecture to
   The Go-CHART has a custom-built battery pack with                           execute the deep learning algorithms using the trained model.
four 18650 batteries that deliver 7.4V at 7000 mAh, which
provides enough power to drive the robot for almost an hour.                   E. Communication capabilities
The battery pack circuit is composed of four batteries, a 2S
battery management system (BMS), the battery protection                           Experiments on cooperative driving and multi-robot con-
circuit, voltage regulators, a buck converter, fuses, and 3.5                  trol will often require ad hoc communication between pairs
mm barrel connectors.                                                          of Go-CHARTs. To implement this communication, the
                                                                               RPi3B onboard the Go-CHART can act as a Wifi router node,
D. Computational capabilities                                                  enabling limited-bandwidth, two-way information transfer
   The Go-CHART is equipped with commercially available                        with a nearby robot. Since this is a short-range weak signal
microprocessors and microcontrollers that are supported by                     with limited bandwidth, usage of the external GPU is not
extensive online resources. High-level control and front-view                  possible (higher bandwidth is required for video streaming
camera video streaming are performed by one Raspberry Pi                       with minimal latency), which prevents the implementation
3 Model B (RPi3B) computer (1.2 GHz quad-core ARM                              of deep learning algorithms that require the GPU. The Go-
Cortex-A53 with 1GB LPDDR2) onboard the robot and                              CHART can also operate independently, without communi-
three RPi Zero W computers (1 GHz single-core CPU, 512                         cating with other Go-CHARTs. In this case, the robot must
MB RAM) dedicated for video streaming from the side and                        have access to the external GPU in order to run controllers
rear-view cameras. The RPi3B runs Ubuntu MATE with the                         that utilize deep learning algorithms.

                                                                          2267

             Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR
In addition, multiple Go-CHARTs can be connected to a                     socket communication protocol to the RPi3B onboard the
common dual band router (multiple routers in the case of                     Go-CHART, which controls the Teensy board to steer the
a large testbed environment), which are in turn connected                    robot in real-time (see next section). A demonstration of
to the same LAN network and operate at different channels                    driving the Go-CHART in Remote-Control mode is shown in
between 2.4 GHz and 5 GHz so as to reduce interference                       the supplementary video, which can also be viewed at [24].
between the routers and increase bandwidth to accommodate
multiple video streams from the Go-CHARTs. The external                      B. Control architecture
Jetson TX2 GPU (one for each Go-CHART) is manually                              The control architecture of the Go-CHART is illustrated
connected to the routers through the LAN to reduce wireless                  in Fig. 5. The RPi3B and Teensy boards transfer informa-
network traffic, and the video stream is processed on the                    tion wirelessly using a two-way USB serial communication
external GPU.                                                                protocol. The Teensy board controls the two motor drivers,
                                                                             each of which drives two motors. All sensors are connected
F. Onboard sensors                                                           to the Teensy board, which transmits the sensor data in real-
   The Go-CHART integrates a variety of vision and range                     time to the RPi3B. The steering angle values obtained by the
sensors, shown in Fig. 4, to enable lane tracking, object                    external computer, either from the lane-tracking algorithm in
detection, collision avoidance, and mapping. There are four                  Autonomous mode or the steering wheel inputs in Remote-
RPi cameras (Sony IMX219 8-megapixel sensor) on the                          Control mode, are sent wirelessly to the RPi3B, which
robot: one ultra-wide angle front-view camera with night                     transmits the commands to the Teensy board to drive the
vision, one camera on each side of the robot, and one rear-                  motors. The RPi3B wirelessly transmits sensor data in real-
view camera. The front-view camera, which has a viewing                      time, including the camera feeds, to the external GPU for
angle of 160◦ , is mounted 9 cm above the bottom of the                      processing. The GPU runs the deep learning algorithms
Go-CHART main body and is tilted downward by an angle                        in Section V for lane tracking and object detection. The
of 30◦ . This placement was chosen so that the camera’s field                GPU then reports the commanded Go-CHART heading and
of view contains the entirety of a 30-cm-tall traffic light                  identified objects of interest in the environment to the RPi3B
when the robot is at least 20 cm from the light, so that                     for further processing. This is done by employing ROS
the robot can still detect the light when it drives forward                  running on the RPi3B with ROS publishers and subscribers.
about one-third of the way into an intersection on the testbed.
An HC-SR04 ultrasonic distance sensor with a range of                               V. NAVIGATION AND O BJECT I DENTIFICATION
2–400 cm and a resolution of 0.3 cm is mounted on the                           This section describes the algorithms used by the Go-
front of the robot, and a 2D 360◦ LiDAR (RPLIDAR A1)                         CHART for lane tracking and object detection. The outer
is mounted on top. The measurements from the ultrasonic                      control loop of the robot performs line and obstacle detection
sensor and LiDAR are used to detect nearby obstacles and                     using techniques from the OpenCV library. The 2D LiDAR
initiate collision avoidance maneuvers, and the LiDAR also                   and ultrasonic sensors detect obstacles, and if an obstacle is
enables 2D mapping of the local environment. In addition,                    detected, the Go-CHART stops immediately and yields until
the Go-CHART includes an LDR (Light Dependent Resistor)                      the obstacle moves away. The robot also detects lines on
sensor for brightness correction of the front-view camera                    the ground in front of it by defining a region of interest in
images, an IMU (Inertial Measurement Unit) with a 3D                         its front-view camera image and performing color masking
accelerometer, 3D compass, and 3D gyroscope, and Hall                        and blob detection. If a line is detected, the robot moves
effect encoders on each motor for odometry.                                  according to the class of a detected object nearby, such as a
                                                                             right-only sign or a stop sign.
  IV. D RIVING M ODES AND C ONTROL A RCHITECTURE
A. Driving modes                                                             A. Lane tracking algorithms
   The Go-CHART can drive in two modes: Autonomous and                          The Go-CHART can use two methods for lane tracking:
Remote-Control. In Autonomous mode, the robot navigates                      (1) techniques from the OpenCV library [16] that apply
through the testbed environment without any human inter-                     Gaussian blur, canny edge detection, color masking, and
vention by tracking lanes and recognizing and responding to                  the Hough transform to a region of interest in the front-
particular objects that it detects with its sensors. In Remote-              view camera image; or (2) end-to-end behavioral cloning
Control mode, a user remotely drives the Go-CHART around                     (NVIDIA neural network model) [12], which imitates human
the testbed using the driving station shown in Fig. 6, which                 driving behaviors. The second method is more computation-
includes a Logitech G920 steering wheel with an accelerator                  ally intensive than the first, but more robust to environmental
and brake pedals. The monitor in the driving station displays                changes if it is trained with sufficiently diverse data sets.
the video feeds from all four cameras on the Go-CHART,                          In the first method, each camera on the Go-CHART is
providing a nearly 360◦ view of the environment in real-                     initially calibrated individually to obtain its intrinsic and
time with low latency (as low as 50 µs). The monitor                         extrinsic parameters. In addition, LDR sensor readings are
also displays odometry readings, which are scaled down                       obtained to perform brightness correction, making the sys-
to mph. The deviations of the steering wheel and pedals                      tem robust to changing lighting conditions (particularly a
from their neutral positions are transmitted remotely using                  change in brightness, which affects OpenCV edge detection

                                                                        2268

           Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
Go-CHART: A miniature remotely accessible self-driving car robot - NSF-PAR
Fig. 4.   Vision and range sensors on the Go-CHART.

                                                      Fig. 5.   Go-CHART control architecture.

techniques). The image obtained from the front-view camera                     In the second method, a deep learning technique called
is converted to gray-scale, and a Gaussian blur of kernel size               end-to-end behavioral cloning (NVIDIA model) is used to
5×5 is performed on a trapezium-shaped region of interest,                   perform lane tracking. The Go-CHART is remotely driven
defined on the lower half of the image. Canny edge detection                 around the testbed, and the front-view camera images and
is applied to this region to detect sharp changes in adjacent                corresponding steering wheel error values are recorded. The
pixel intensity. Following this step, the Hough transform is                 neural network is trained with these error values as inputs.
performed to accurately detect the boundaries of the lanes                   Once trained, the network can replicate the same input
in the image, which are used to compute the steering angle                   behavior exhibited by the driver. Then, for an input camera
required for the Go-CHART to stay within the lanes.                          image, the network outputs an error value that is used in the

                                                                        2269

           Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
VI. E XPERIMENTS
                                                                                  In this section, we demonstrate the capabilities of the Go-
                                                                               CHART on the small-scale driving testbed shown in Fig. 1
                                                                               and evaluate the performance of its lane tracking and object
                                                                               identification algorithms.
                                                                               A. Lane tracking
                                                                                  Demonstrations of the two lane-tracking methods de-
                                                                               scribed in Section V-A are shown in the supplementary
                                                                               video. When the Go-CHART uses the first lane-tracking
                                                                               method, the blue lines in the video indicate the lane sections
                                                                               that the robot tracks. This method performs lane tracking
                                                                               in real-time at 30 FPS without any latency. To implement
                                                                               the second lane-tracking method, the neural network was
                                                                               trained as described in Section V-A. Each camera image in
                                                                               the training data set was preprocessed as follows in order
                                                                               to meet the requirements of the NVIDIA neural network.
                                                                               First, a trapezium-shaped region of interest is defined in
                                                                               the image, and this region is converted to YUV color
                                                                               space. The trapezium image is stretched and resized into a
                                                                               rectangular image. Three regions of dimensions 352 × 288
                                                                               pixels are selected from this rectangular image and then are
                                                                               translated by 25 pixels and rotated by 1◦ with respect to
                                                                               each other. Around 3000 images and steering error values
                                                                               were collected; 80% were utilized as training data, and
                                                                               the remaining 20% were used as validation data. Since the
                                                                               available data set was limited, the images were subjected
                                                                               to data augmentation techniques to improve learning and
                                                                               significantly reduce the training loss. Using the RTX 2080 Ti
                                                                               GPU, the neural network required around 8 min to train per
                                                                               epoch. The network was trained for a total of 300 epochs.
                                                                               The trained model can be used to predict the Go-CHART
                                                                               steering angle for the given input image using the external
Fig. 6.    Top: Driving station setup for Remote-Control driving mode.         GPU. When implemented on the Jetson TX2, the network
Bottom: Monitor in the driving station displaying video streams from the       predicts the steering angle at a rate of approximately 5-7
front-view, side-view, and rear-view cameras onboard the Go-CHART.             angle values per second for the given input image size.
                                                                               B. Object detection
PD steering controller (see Section III-B).                                       In order to run both the YOLOv3 and Tiny YOLO
                                                                               object detection algorithms, the video feed from the front-
                                                                               view camera (frame size 320 × 240 pixels) was wirelessly
B. Object detection algorithms                                                 streamed at 30 FPS over the wireless network to the GPU
                                                                               on which the trained neural network model is executed to
   To enable the Go-CHART to perform real-time detection                       predict the object classes of interest.
of objects in the testbed, such as other Go-CHARTs and                            To detect U.S. traffic signs and signals and other Go-
traffic signs and signals, we tested YOLOv3 [10], Tiny                         CHARTs, we needed to create a custom data set of images.
YOLO [11], ResNet [15], and Faster RCNN [13]. Ultimately,                      Both YOLOv3 and Tiny YOLO were trained on custom
YOLOv3 [10] and Tiny YOLO [11] were selected since they                        data by transfer learning from the pre-trained COCO data
detected objects with high accuracy at considerably higher                     set [14], which consists of around 200,000 labeled images
frame rates than the other algorithms, given limited GPU                       and around 91 different object classes. We trained the neural
processing power (128 CUDA cores), and performed well                          network on 10 custom object classes, shown at the bottom of
on low-resolution live video streams with varying latency.                     Fig. 7: speed limit sign, stop sign, pedestrian crossing sign,
This enabled their use for real-time object detection, which                   traffic light sign, traffic signal (red, yellow, and green), T
was critical for the Autonomous driving mode. Low detection                    intersection sign, turning vehicles yield sign, and Go-CHART
rates and low identification accuracy can cause the robot                      robot. The custom data set was collected from the front-
to miss critical environmental features, potentially causing                   view camera wireless video stream, stored in an external
accidents.                                                                     computer, and processed using techniques such as resizing

                                                                          2270

             Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
Fig. 7. YOLOv3 object detection results for the video feed from the Go-CHART front-view camera, with bounding boxes and class confidence levels
shown. The neural network was trained on the object classes given in the bottom row. Note: The traffic signal includes 3 different object classes: red,
yellow and green.

                                                                                                           TABLE III
and annotation. It is important to record the images from the
                                                                                 I NFERENCE SPEED AND MEAN AVERAGE PRECISION OF YOLOV 3             AND
wireless stream, although some images might be blurred and
                                                                                               T INY YOLO, TESTED ON J ETSON TX2
low-resolution, so that the network is trained to be robust to
such images. This eases the strain on the network bandwidth                        Algorithm      Input image      Inference          mAP (Mean Av-
due to video streams from multiple Go-CHARTs, and also                                            size in pixels   speed (FPS)        erage Precision)
enables the YOLO network to produce faster results. Around                                        640 × 480         0.8               76.39
                                                                                                  480 × 360         1.9               75.37
300 different images per class were obtained (a total of                           YOLOv3
                                                                                                  320 × 240         4.7               68.96
around 3000 images), and bounding boxes were manually                                             160 × 120         9.1               22.21
marked on the regions of interest in all the images with their                                    640 × 480         8.3               54.62
                                                                                                  480 × 360         12.6              41.72
corresponding class labels. The training took around 30 min                        Tiny YOLO
                                                                                                  320 × 240         18.3              34.38
per epoch and required at least 20 epochs to obtain accurate                                      160 × 120         19.9              16.79
predictions. However, we trained for 100 epochs in order
to obtain higher confidence values for the detected classes.
The number of epochs required for training decreases as the
                                                                                  Figure 7 and the supplementary video show object detec-
number of classes and size of the training data set for each
                                                                               tion results using the YOLOv3 algorithm. If one or multiple
class increase.
                                                                               object classes of interest are present in a particular frame
  After the model was trained, it was executed on the Jetson                   from the video feed, the algorithm displays the detected
TX2. The model is loaded after an initial boot-up delay                        object class(es), the prediction accuracy for each class, and
of up to 90 s. Once loaded, object detection is performed                      the bounding box dimension and position of the bounding
considerably quickly, at around 3-5 FPS for YOLOv3 and                         box. The prediction accuracy is set at a threshold value of
up to 18 FPS for Tiny YOLO. The major trade-off between                        40% to ignore false positives. This threshold is relatively low
the two algorithms is that YOLOv3 provides accurate results                    because the prediction accuracy can decrease significantly for
but requires higher GPU usage, while the opposite is true for                  blurry, low-resolution images from the camera on a moving
Tiny YOLO. Table III lists the inference speeds and mAP                        Go-CHART. Despite this, Figure 7 and the supplementary
(Mean Average Precision) of both the YOLOv3 and Tiny                           video show that the algorithm is generally effective at
YOLO algorithms, tested for four image sizes with 10 input                     identifying objects in the testbed.
images of each size.                                                              The Go-CHART decides on its next action based on the

                                                                          2271

             Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
type(s) of objects that it currently detects and the relevant                   [5] Sertac Karaman, Ariel Anders, Michael Boulet, Jane Connor, Kenneth
traffic rules (e.g., stopping when a red light or stop sign                         Gregson, Winter Guerra, Owen Guldner, et al. Project-based, collab-
                                                                                    orative, algorithmic robotics for high school students: Programming
is detected). When the Go-CHART detects multiple traffic                            self-driving race cars at MIT. In 2017 IEEE Integrated STEM Educa-
lights at an intersection, it bases its next action on the traffic                  tion Conference (ISEC), pp. 195-203. IEEE, 2017.
light with the largest bounding box, since that light is likely                 [6] Daniel Pickem, Paul Glotfelter, Li Wang, Mark Mote, Aaron Ames,
                                                                                    Eric Feron, and Magnus Egerstedt. The Robotarium: A remotely
to be the closest one to the robot. The supplementary video                         accessible swarm robotics research testbed. In 2017 IEEE Interna-
shows scenarios in which (1) the Go-CHART stops at a red                            tional Conference on Robotics and Automation (ICRA), pp. 1699-1706.
light in an intersection and resumes driving when the light                         IEEE, 2017.
                                                                                [7] Joseph Betthauser, Daniel Benavides, Jeff Schornick, Neal O’Hara,
turns green, and (2) the robot stops at a stop sign and waits                       Jimit Patel, Jeremy Cole, and Edgar Lobaton. WolfBot: A distributed
for another Go-CHART with the right-of-way to cross in                              mobile sensing platform for research and education. In Proceedings of
front of it before driving forward.                                                 the 2014 Zone 1 Conference of the American Society for Engineering
                                                                                    Education, pp. 1-8. IEEE, 2014.
                                                                                [8] Adam Stager, Luke Bhan, Andreas Malikopoulos, and Liuhui Zhao.
       VII. CONCLUSION AND FUTURE WORK                                              A scaled smart city for experimental validation of connected and
                                                                                    automated vehicles. arXiv preprint arXiv:1710.11408 (2017).
   In this paper, we have presented the design, capabili-                       [9] Rakshith Subramanyam (2018). CHARTOPOLIS: A Self Driving Car
                                                                                    Test Bed. Master’s Thesis in Electrical Engineering, Arizona State
ties, and control architecture of the Go-CHART, a low-cost                          University.
miniature robot that emulates a self-driving car. The Go-                      [10] Joseph Redmon and Ali Farhadi. YOLOv3: An incremental improve-
CHART can also be controlled remotely by a user who views                           ment. arXiv preprint arXiv:1804.02767 (2018).
                                                                               [11] Rachel Huang, Jonathan Pedoeem, and Cuixian Chen. YOLO-LITE: A
the environment through the robot’s four onboard cameras                            real-time object detection algorithm optimized for non-GPU comput-
from a driving station. This capability will enable us to                           ers. In 2018 IEEE International Conference on Big Data (Big Data),
conduct experiments on interactions between human-driven                            pp. 2503-2510. IEEE, 2018.
                                                                               [12] Mariusz Bojarski, Davide Del Testa, Daniel Dworakowski, Bernhard
and self-driving vehicles in a safe, controlled environment.                        Firner, Beat Flepp, Prasoon Goyal, Lawrence D. Jackel et al. End to
We describe lane-tracking and object detection algorithms                           end learning for self-driving cars. arXiv preprint arXiv:1604.07316
and demonstrate their implementation on a Go-CHART that                             (2016).
                                                                               [13] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster
autonomously navigates a miniature driving testbed, with its                        R-CNN: Towards real-time object detection with region proposal
processing power augmented by an external GPU. We plan to                           networks. In Advances in Neural Information Processing Systems, pp.
expand the driving testbed to include additional traffic signs,                     91-99. 2015.
                                                                               [14] Xinlei Chen, Hao Fang, Tsung-Yi Lin, Ramakrishna Vedantam,
dynamic lighting conditions, reconfigurable buildings, and                          Saurabh Gupta, Piotr Dollár, and C. Lawrence Zitnick. Microsoft
media projections to customize the environment. We also                             COCO captions: Data collection and evaluation server. arXiv preprint
plan to add artificial fog, which will enable us to develop                         arXiv:1504.00325 (2015).
                                                                               [15] Sasha Targ, Diogo Almeida, and Kevin Lyman. Resnet in Resnet:
algorithms for real-time defogging using only the robot’s on-                       Generalizing residual architectures. arXiv preprint arXiv:1603.08029
board monocular cameras. In addition, we will replace the                           (2016).
robot’s front-view camera with an RGBD camera to enable                        [16] G. Bradski. The OpenCV Library. Dr. Dobb’s Journal of Software
                                                                                    Tools, vol. 25, no. 11, pp. 120–126, 2000.
tests of real-time visual SLAM algorithms.                                     [17] Google Maps: Miniatur Wunderland. Accessed: 2020-07-31.
                                                                                    [Online] Available: https://www.google.com/maps/about/behind-the-
                   ACKNOWLEDGMENTS                                                  scenes/streetview/treks/miniatur-wunderland/
                                                                               [18] University of Michigan Mcity. Accessed: 2020-07-31. [Online] Avail-
   The authors thank Sreenithy Chandran (ASU) for helping                           able: https://mcity.umich.edu/
                                                                               [19] Karen Hao. “The latest fake town built for self-driving cars has
to implement the deep learning techniques and Sangeet Ulhas                         opened in South Korea.” Quartz, Nov. 6, 2017. Accessed: 2020-07-31.
(ASU), Rakshith Subramanyam (ASU), Karthik Ganesan                                  [Online] Available: https://qz.com/1121372/south-korea-opens-k-city-
(ASU), and Sritanay Vedartham (BASIS Scottsdale) for                                the-latest-fake-town-built-for-self-driving-cars/
                                                                               [20] Millbrook:        Connected         and       Autonomous       Vehicle
helping build the testbed.                                                          Testing.      Accessed:       2020-07-31.        [Online]   Available:
                                                                                    https://www.millbrook.us/services/connected-and-autonomous-
                            R EFERENCES                                             vehicle-testing/
                                                                               [21] Srinivasa, S.S., Lancaster, P., Michalove, J., Schmittle, M., Rockett,
                                                                                    C.S.M., Smith, J.R., Choudhury, S., Mavrogiannis, C. and Sadeghi, F.
 [1] Liam Paull, Jacopo Tani, Heejin Ahn, Javier Alonso-Mora, Luca
                                                                                    MuSHR: A low-cost, open-source robotic racecar for education and
     Carlone, Michal Cap, Yu Fan Chen, et al. Duckietown: an open, in-
                                                                                    research. arXiv preprint arXiv:1908.08031 (2019).
     expensive and flexible platform for autonomy education and research.
                                                                               [22] Balaji, B., Mallya, S., Genc, S., Gupta, S., Dirac, L., Khare, V.,
     In 2017 IEEE International Conference on Robotics and Automation
                                                                                    Roy, G., Sun, T., Tao, Y., Townsend, B. and Calleja, E. DeepRacer:
     (ICRA), pp. 1497-1504. IEEE, 2017.
                                                                                    Educational autonomous racing platform for experimentation with
 [2] Sean Wilson, Ruben Gameros, Michael Sheely, Matthew Lin, Kathryn               sim2real reinforcement learning. arXiv preprint arXiv:1911.01562
     Dover, Robert Gevorkyan, Matt Haberland, Andrea Bertozzi, and                  (2019).
     Spring Berman. Pheeno, a versatile swarm robotic research and             [23] W. Roscoe. Donkey Car: An opensource DIY self driving plat-
     education platform. IEEE Robotics and Automation Letters vol. 1, no.           form for small scale cars. Accessed: 2020-07-31. [Online] Available:
     2, pp. 884-891, July 2016.                                                     https://www.donkeycar.com
 [3] Nicholas Hyldmar, Yijun He, and Amanda Prorok. A fleet of miniature       [24] NVIDIA        Autonomous       Machines.      Accessed:   2020-07-31.
     cars for experiments in cooperative driving. In 2019 IEEE Interna-             [Online]      Available:    https://www.nvidia.com/en-us/autonomous-
     tional Conference on Robotics and Automation (ICRA), pp. 3238-3244.            machines/embedded-systems/jetbot-ai-robot-kit/
     IEEE, 2019.                                                               [25] “Go-CHART: A miniature remotely accessible self-driving car
 [4] Alexander Liniger, Alexander Domahidi, and Manfred Morari.                     robot,” Autonomous Collective Systems Laboratory Youtube channel,
     Optimization-based autonomous racing of 1:43 scale RC cars. Optimal            https://www.youtube.com/watch?v=pAa61VpF6oQ
     Control Applications and Methods 36(5):628-647, 2015.

                                                                          2272

             Authorized licensed use limited to: ASU Library. Downloaded on May 18,2021 at 05:13:24 UTC from IEEE Xplore. Restrictions apply.
You can also read